A computational framework for confidentiality-aware data transformation and decision-preserving Taguchi optimization

Industrial optimization frequently requires collaborative analysis when original production responses cannot be disclosed. This study investigates whether response transformations retain Taguchi optimization decisions under confidentiality constraints. Parabolic, scaled-logit (SLT), and Yeo–Johnson transformations were compared using Taguchi cross-array simulations. A DOE-structured calibration–validation design separated transformation-rule development from held-out assessment, with rules fixed before generating fresh validation responses. Preservation was assessed through factor and treatment rankings, top-factor and recommended-setting agreement, and signal-to-noise (S/N) and mean-response patterns. Numerical masking and reconstruction were evaluated separately. Under nominal-the-best conditions, Yeo–Johnson preserved exact factor rankings in 92.19% of held-out evaluations and recommended settings in 93.75%. Corresponding rates were 72.66% and 78.91% for SLT, versus 53.91% and 68.75% for parabolic transformation. Extended Monte Carlo analysis yielded 91.25% exact ranking agreement for Yeo–Johnson. However, numerical masking was limited: Yeo–Johnson and SLT allowed recovery to numerical precision when complete transformation specifications were known. An HDD pivot-manufacturing case further examined Yeo–Johnson using measured responses. Under noise-inclusive conditions, original and transformed analyses produced identical factor rankings, treatment rankings, recommended settings, and best observed treatments. Compared with the reference treatment, the selected robust treatment showed 48.14% lower response standard deviation and 44.79% lower target-centered RMSE within the experimental data. These comparisons reflect observed differences within the experimental dataset and do not constitute independent confirmation of improvements after implementation in routine production. These findings support evaluating transformations by both downstream decision preservation and numerical masking. The framework enables structured assessment of analytical sharing under confidentiality constraints without claiming formal privacy or cryptographic protection.